Traditional market research delivers valuable insights, but what happens when you want to explore scenarios beyond your original study scope? Enter "study boosting", a synthetic research method that bridges real human experiences with AI personas.

The Challenge: From Static Reports to Dynamic Research

Most research projects end with a final report. But what if stakeholders want to test new messaging, explore different scenarios, or validate concepts that weren't part of the original study? Traditionally, this meant commissioning entirely new research, which is expensive and time-consuming. 

Study boosting offers a different approach: transform your existing survey respondents into AI personas that retain their authentic characteristics, then use these personas for ongoing research simulation.

Our Process: From Survey Data to Validated AI Personas

We recently completed a study boosting project for automotive market research across Germany, Netherlands, and Portugal. Here's how we transformed real survey responses into production-ready AI personas.

Step 1: Persona Generation from Survey Segments

Starting with the client's raw survey data and segmentation model, we organized responses by segment and generated AI personas using our clone-via-text feature (available on our Smart plan at $100/month). Each persona authentically reflected the traits and characteristics evident in their segment's survey responses.

This automotive study spanned three European markets across three purchase funnel stages, creating nine distinct audiences like this.

Market 1

• [country] / [awareness]

• [country] / [consideration]

• [country] / [action]

Market 2

• [country] / [awareness]

• [country] / [consideration]

• [country] / [action]

Market 3

• [country] / [awareness]

• [country] / [consideration]

• [country] / [action]

Given budget constraints and the client's focus on exploratory message testing for new localised product launches, we used straightforward random sampling from raw responses. While this approach effectively supported concept exploration and brand association discovery, there's potential to enhance these audiences for quantitative study boosting applications. (DM me if you’re interested about that).

Each persona received comprehensive demographic profiles including background, occupation, and lifestyle details aligned with their segment characteristics. Most importantly, we embedded their actual survey responses as authentic memories, ensuring each AI persona carried the genuine voice and preferences of real respondents rather than generic demographic assumptions.

Step 2: Memory Enrichment with Individual Survey Responses

This step was critical. I enriched each AI persona with their actual survey responses as "memories." by appending the survey data to the Persona Background Bio. Just as I enriched the persona below using memories from tweets, but here I’ve incorporated the entire survey along with authentic human responses.



This was a strategic bet to try and ensure our AI personas don’t just represent demographic archetypes, but they carry the authentic voice and preferences of real respondents. An audience design tactic that worked well for me in a PMF prediction experiment.

In an ideal world, we’d have had the budget to run our calibration engine that simply removes the guesswork at this stage (and bias), by running thousands of experiments to see what actually results in accurate study replication. But our champion and I treated this as a paid POC to warm stakeholders and make everyone aware of the context engineering effort required for manual experiments.

Step 3: Data Quality Validation

Accuracy transparency is core to how Mike and I run AskRally, so I checked how well we were matching real survey data to cloned AI persona. That’s when I discovered a significant data quality issue where the wrong survey responses were being enriched to the wrong persona, likely stemming from an unexpected async workflow in our bulk CSV clone.

The challenge with this survey-based memory enrichment step is that it requires an advanced, multi-layered personification process. It’s not as simple as checking whether “female in survey = female in AI persona.” Some survey responses were free-text, which introduces ambiguity and multiple possible mismatches between the survey and the persona. So I built custom validation scripts using an LLM-as-judge, that compared persona backgrounds with their survey memories, checking for discrepancies accross:

• Demographics (age, gender)
• Income levels  
• {redacted product} preferences
• Lifestyle factors
• Purchase timelines
• Brand preferences

It worked really well on first try and helped flag problematic personas. 

Step 4: Data Cleaning and Manual Enrichment

For corrupted audiences, we had two options given the constraints at the time:

1. Remove problematic personas - Keep only validated matches

2. Manual re-enrichment - Carefully match personas with appropriate survey responses

We reran the key markets and high error rate audiences and for the others just dropped the problem personas until arriving at the following audience accuracies. 

Cleanup results:

• Market 1 Awareness:        96.4% 
• Market 1 Consideration:  96.8%  
• Market 1 Action:               100%   
• Market 2 Awareness:       100%  
• Market 2 Consideration: 100%   
• Market 2 Action:              100% 
• Market 3 Awareness:       100% 
• Market 3 Consideration:  100% 
• Market 3 Action:               94.4% 

Step 5: Localization for Regional Testing

With message testing as our primary use case, we made a bet on cultural authenticity. We wanted to try and avoid a common AI bias: having personas respond from an English-speaking perspective when evaluating local market content.

The challenge: For genuine localization insights, AI personas must think and respond as native speakers would, not through the lens of translated English responses. This distinction is crucial when testing regional messaging resonance and cultural relevance.

Our approach: We translated personas into their native languages while maintaining complete survey data integrity:

The Deliverable: Research-Ready AI Personas

The final output: a ready-to-survey-boost set of AI persona audiences with:

• near 100% data quality (validated persona-survey alignment)

• authentic memories (real survey responses embedded)

• local language versions (for regional message testing)

• segmented by research objectives (Action, Consideration, Awareness)

Study Boosting in Action: From Static to Simulation

These validated audiences were then loaded into the client’s AskRally account who generated an API key to power our google sheet template. This is how you enable your study consumers (without an AskRally account) to run questions to these personas and get almost instant answers back.

We wanted a soft landing into synthetic testing for a new audiences, so we spruced up the gsheet template with some custom documentation, tidy audience references, and template survey questions ready to copy simulate.  

 Examples of How AI Personas React to Reels

The Future of Hybrid Research

Study boosting represents a new category: hybrid synthetic research that combines the authenticity of traditional methods with the flexibility of AI simulation. Researchers get the best of both worlds, grounded human insights with unlimited scenario testing capability 24/7.

Rather than viewing AI personas as replacements for human research, study boosting positions them as extensions of existing studies, allowing stakeholders to explore new questions using the same validated respondent base.

Want to see study boosting in action or discuss the possibilities with AI persona? Book me for a demo.